[llvm] [InstCombine] Fold `(icmp pred (trunc nuw/nsw X), C)` -> `(icmp pred X, (zext/sext C))` (PR #87935)

[Yingwei Zheng via llvm-commits]

================
@@ -1409,6 +1409,19 @@ Instruction *InstCombinerImpl::foldICmpTruncConstant(ICmpInst &Cmp,
                                                      const APInt &C) {
   ICmpInst::Predicate Pred = Cmp.getPredicate();
   Value *X = Trunc->getOperand(0);
+  Type *SrcTy = X->getType();
+  unsigned DstBits = Trunc->getType()->getScalarSizeInBits(),
+           SrcBits = SrcTy->getScalarSizeInBits();
+
+  // Match (icmp pred (trunc nuw/nsw X), C)
+  // Which we can convert to (icmp pred X, (sext/zext C))
+  if (shouldChangeType(DstBits, SrcBits)) {
+    if (!Cmp.isSigned() && Trunc->hasNoUnsignedWrap())
+      return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, C.zext(SrcBits)));
+    if (Trunc->hasNoSignedWrap())
+      return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, C.sext(SrcBits)));
----------------
dtcxzyw wrote:

I prefer the sign-extended constant as it may be cheaper to materialize than the zero-extended one (e.g., -1 vs 4294967296).

Don't do this until we have some performance data.


https://github.com/llvm/llvm-project/pull/87935